Epic's agents are here: what stays your job
This week at its UGM conference, Epic showed its AI agents working as one system. Art works with clinicians, Emmie with patients, and Penny with the revenue cycle. Epic also featured Agent Factory, its no-code tool for building your own agents, and predictions from Curiosity, its models trained on Cosmos data, which 20 customers are validating.
If you run data or analytics at a health system on Epic, you probably had two reactions. This is impressive. And what does it mean for my team?
Let me be fair to Epic first. Its agents work where clinicians and staff already work. They see the chart in full. Cosmos gives Epic a base of data for training models that few others can match. And its AI is already in wide use: Epic says more than 85% of its customers use Epic AI.
For work that lives in the EHR, like drafting a note or answering a patient message, I’d expect Epic’s agents to be very good. If you’re on Epic, you should use them.
That still leaves real work for your team. Here’s what I think stays yours.
The data outside the EHR
Many of the costliest questions in a health system don’t live in the chart.
Costs are a good example. Kaufman Hall reported this week that supply and drug costs are rising well above inflation. Vizient projects hospital pharmacy prices will rise 3.54% in 2027, and that figure leaves out tariffs. The answers to those cost questions sit in your ERP, your invoices, your contracts and your general ledger.
The same goes for labor. Payroll, timekeeping and scheduling hold the data on what your workforce costs.
Penny can do a lot with the revenue cycle as the EHR sees it. Asking why margin fell on a service line takes more. You need supply costs, labor costs and payer contracts lined up next to the clinical data.
That shared place, where the EHR’s data meets everything else, is still yours to build and keep right. Whether the agents that use it come from Epic or from your own team, they can only be as good as what they can reach.
Measurement
Every vendor will tell you what its agents achieved. Those numbers are useful. They are also the vendor’s numbers.
Your job is to know your own baseline before an agent goes live, and to measure the change after, in terms your CFO trusts. Minutes per note. Days in accounts receivable. Denials overturned. Messages answered without a callback.
Pick the measure before you turn the agent on. Write down today’s number. Check it at 90 days. Watch for anything that got worse, too, like longer waits or more rework somewhere downstream.
The same goes for predictions. It’s a good sign that 20 customers are validating Curiosity. It doesn’t tell you how a model performs on your patients and your workflows. Checking that is your work, and it should be built into how you roll out every prediction.
Governance
Agents take actions. Someone has to decide what each agent may do on its own, what needs a person’s approval, and who watches it over time.
That responsibility doesn’t move to the vendor. The Joint Commission’s voluntary AI certification, launched in June, makes the point. It certifies an organization’s AI governance, monitoring and training, not individual tools. In late July, Hackensack Meridian Health became the first system to earn it.
Agent Factory makes this more pressing. When building an agent takes no code, more people will build them. That’s good for speed. It also means you need a simple list of every agent you run, with an owner, a purpose and a way to check it’s still working.
None of this is a reason to slow down on Epic’s agents. It’s how you get the most from them.
Before your next Epic AI planning meeting, write down every agent you’ve turned on or plan to turn on. Next to each one, name the owner, the measure and the data it reads. Any blank on that page is work that stays with your team.
Discussion